retinal vessel segmentation based on adaptive random · pdf fileretinal vessel segmentation...
TRANSCRIPT
Retinal Vessel Segmentation Based on Adaptive
Random Sampling
Nicolas Jouandeau, Zhi Yan, and Patrick Greussay Advanced Computing Laboratory of Saint-Denis (LIASD), Paris 8 University, 93526 Saint-Denis, France
Email: {n, yz, pg}@ai.univ-paris8.fr
Beiji Zou and Yao Xiang School of Information Science and Engineering, Central South University, Changsha 410083, PR China
Email: {bjzou, yao.xiang}@mail.csu.edu.cn
Abstract—This paper presents a method for the extraction
of blood vessels from fundus images. The proposed method
is an unsupervised learning method which can automatically
segment retinal blood vessels based on an adaptive random
sampling algorithm. This algorithm consists in taking an
adequate number of random samples in fundus images, and
all the samples are contracted to the position of the blood
vessels, then the retinal vessels will be revealed. The basic
algorithm framework is presented in this paper and several
preliminary experiments validate the feasibility and
effectiveness of the proposed method.
Index Terms—
learning, sampling algorithm, retinal vessels, fundus images
I. INTRODUCTION
Retinal vessel segmentation refers to extract blood
vessels in retinal camera images. This technique is of
important significance on screening, diagnosing and
treating various ophthalmological and cardiovascular
diseases. Manual segmentation of retinal blood vessels
can guarantee the accuracy of the segmentation, but it is a
long work which also needs some professional trainings.
The public desires to achieve automation by means of
computers. This paper introduces a new method for
automatic retinal vessel segmentation based on an
adaptive random sampling algorithm. The basic idea is
(see Fig. 1): take a random sample (i.e. a random pixel) p
in the fundus image, if there exists a pixel q around p
which has a lower value (i.e. the color of q is darker than
that of p), then p retracts to q and marked as vessel. This
method is actually to find the local optimal solutions, and
then combined into a global optimal solution.
Figure 1. An example illustrating the sampling-based method.
Manuscript received May 21, 2013; revised August 10, 2013.
The rest of the paper is organized as follows: Section II
describes an overview of some related works; Section III
describes our adaptive random sampling algorithm;
Section IV presents the experimental results obtained
with the proposed method; and the paper is concluded in
Section V at last.
II. RELATED WORK
Retinal vessel segmentation is actually an issue of
image processing, which has been widely studied since
the late 1980s. Existing methods can be generally divided
into the following categories [1]: 1) machine learning
(mainly including supervised learning and unsupervised
learning), 2) matched filtering, 3) vessel tracking, 4)
morphological processing, 5) multi-scale approaches, 6)
model based approaches, and 7) parallel hardware based
approaches. Experiments show that the accuracy of
segmentation can reach up to 97%.
However, among the existing works, the development
of unsupervised learning methods is relatively small, and
the accuracy of segmentation by using these methods is
relatively low. [2] presented an unsupervised method for
segmenting the vessel from color retinal images. The
vessels are modeled as trenches and the medial lines of
the trenches are extracted using the curvature information
derived from a novel curvature estimate. The output of
this process is in the form of a medial axes map of the
vessels. The entire vessel structure is then extracted using
a region growing method. Their method achieves an area
under the ROC (Receiver Operating Characteristic) curve
of 0.9271 and a maximum average accuracy of 0.9361 on
the publicly available DRIVE database. [3] presented
another unsupervised method by extracting the intensity
information from red and green channels of the same
retinal image to correct non-uniform illumination in color
fundus images. Matched filtering is utilized to enhance
the contrast of blood vessels against the background. The
enhanced blood vessels are then segmented by employing
spatially weighted fuzzy c-means clustering. The
evaluation of the method resulted in an area under the
ROC curve of 0.9518 and 0.9298, and an average
accuracy of 0.8911 and 0.8976, respectively on the
DRIVE and STARE databases.
Journal of Medical and Bioengineering Vol. 3, No. 3, September 2014
199©2014 Engineering and Technology Publishingdoi: 10.12720/jomb.3.3.199-202
blood vessel segmentation, unsupervised
Our research focuses on the unsupervised method for
its full autonomy. By using the unsupervised method, we
don't need any training sample and manual intervention
(e.g. manual labelling). Moreover, our research focuses
on the possibility of solving complex problems with
simple methods, which can reduce the dependence on
hardware, so it has more practical significance.
III. OUR METHOD
Our proposed method is inspired by the sampling-
based methods in the field of robotics [4]. Sampling-
based methods are currently considered state of the art for
robot motion planning in high-dimensional spaces, and
have been highly successful in solving many complex
problems which have dozens or even hundreds of
dimensions. These methods can avoid the problem of
local minima, but they are probabilistically complete,
meaning they cannot guarantee to find a solution if one
exists. The fundus image could be considered to be a
complex environment. Using a probabilistic method can
meet the requirement of high efficiency and low
consumption. Our method for retinal vessel segmentation
by using the adaptive random sampling algorithm is
illustrated in Algorithm 1.
Algorithm 1 Adaptive Random Sampling
1: while a maximum number of pixels n pixels has
been marked as blood vessel do
2: for number of sample i = 0 to maximum
number m do
3: generate a random sample p in retinal image
4: if exists q in adjacent area alpha of p : |p -
q| > threshold then
5: if q < p then
6: q is labeled as vessel
7: else
8: p is labeled as vessel
9: end if
10: end if
11: end for
12: end while
In Algorithm1, line 1 means that the program will be
repeated until convergence has been reached. Line 2
indicates that we must take a sufficient number (denoted
by m) of sample pixels. In line 4, |p-q| indicates the
difference between the two pixel values p and q. Line 6
can actually be regarded as the retraction of the sample p
to the location of the pixel q. That is the reason why we
call this algorithm the adaptive random sampling. In the
current version of the algorithm, n, m and threshold are
three variables that have been assigned a value at the
beginning of the program. This manner of manual setting
must be difficult to deal with a large number of images. It
will be changed to an automatic manner in the program
according to various images in our future
implementations.
IV. EXPERIMENTS
A. Experimental Setup
Several primary experiments have been conducted to
evaluate our proposed method. All experiments reported
in this paper were carried out on the Ubuntu 10.04
operating system with an Intel Core 2 Duo T5500
1.66GHz processor, an Intel 945GM Express chipset and
two DDR2 800MHz 512MB dual channel memory. Our
program is realized with secondary development tools
based on libpng library, which can be downloaded from
www.ai.univ-paris8.fr/~yz/ARS.tar.gz.
B. Experimental Results
The first stage of the experiment is to transfer a color
fundus image to a target image. This step is essential to
vessel segmentation because it is meaningful to find an
effective objective metric [5]. In our method, the color
fundus image is converted to a magenta channel image in
CMYK (Cyan, Magenta, Yellow and blacK) color model.
In contrast with the green channel image in RGB (Red,
Green and Blue) color model, greyscale image, and color
image, retinal vessels are more distinct in the magenta
channel image in CMYK color model [6]. The magenta
channel image is also more suitable for our adaptive
random sampling algorithm, by reason of the obvious
difference between vessels and other areas, thus more
sensitive to the threshold. A visual comparison of the
different images is given in Fig. 2.
(a) Original color image
(b) Greyscale image
(c) Green channel image (RGB color model)
Journal of Medical and Bioengineering Vol. 3, No. 3, September 2014
200©2014 Engineering and Technology Publishing
(d) Magenta channel image (CMYK color model)
Figure 2. A visual comparison of different models of a fundus image.
The experimental fundus image is selected from the
ImageRet database [7]. Fig. 3 shows the experimental
results for four critical areas in the image. These four
areas are sequentially labeled as area (A), area (B), area
(C) and area (D), in accordance with the clockwise
direction. Area (A) represents the vessel branch and cross,
area (B) represents the blood vessels on optic disc, area
(C) represents the central vessel reflex, and area (D)
represents the blood capillaries. As can be seen from Fig.
3, the current version of our method can identify vessel
branch and cross, vessels on optic disc, and central vessel
reflex effectively, but it cannot handle blood capillaries
very well, because the pixel values of the capillary are too
close to the values of the other area. This problem will be
improved in our future research.
Figure 3. Four critical areas in a fundus image.
Fig. 4 depicts the vessel extracted by using our
sampling-based method. The left image (subfigure (a))
represents the initial result. The vessels can be obviously
identified from this image, but it still has a lot of noises
for reasons of uneven background, macular and exudates.
The right image (subfigure (b)) represents the result after
preliminary noise reduction. In fact, the noises produced
by using the current version of our method are mainly
grouped around the macular and the exudates. From
another point of view, it can be used in detection of
pathology.
(a) Initial result
(b) After preliminary noise reduction
Figure 4. Vessel extracted by using the sampling-based method.
V. CONCLUSION
In this paper, we presented a new retinal vessel
segmentation method based on an adaptive random
sampling algorithm. The idea is to take an adequate
number of random samples in retinal camera images, and
all the samples should retract to the location of the vessel
so as to present it. A basic algorithm framework and
Journal of Medical and Bioengineering Vol. 3, No. 3, September 2014
201©2014 Engineering and Technology Publishing
some preliminary experimental results are presented. The
experimental results proved the feasibility of our
proposed method. The future work will focus on:
Algorithm need to be improved, especially the
determination of the values of n, m and threshold.
Image noise should be reduced effectively.
More images need to be tested, and experimental
results should be quantified with accuracy.
ACKNOWLEDGMENT
This work has partly been supported by the Sino-
French scientific project “Xu Guangqi” under contract
number PHC 28011WC.
REFERENCES
[1] M. Fraz, P. Remagnino, A. Hoppe, B. Uyyanonvara, A. Rudnicka,
C. Owen, and S. Barman, “Blood vessel segmentation
methodologies in retinal images - a survey,” Computer Methods and Programs in Biomedicine, vol. 108, no. 1, pp. 407-433,
October 2012.
[2] S. Garg, J. Sivaswamy, and S. Chandra, “Unsupervised curvature-based retinal vessel segmentation,” in Proc. 4th IEEE
International Symposium on Biomedical Imaging, Washington DC, USA, April 2007, pp. 344-347.
[3] G. B. Kande, P. V. Subbaiah, and T. S. Savithri, “Unsupervised
fuzzy based vessel segmentation in pathological digital fundus images,” Journal of Medical Systems, vol. 34, no. 5, pp. 849-858,
October 2010. [4] Z. Yan, N. Jouandeau, and A. Ali Cherif, “ACS-PRM: Adaptive
cross sampling based probabilistic roadmap for multi-robot motion
planning,” in Proc. 12th International Conference on Intelligent Autonomous Systems, Jeju Island, Korea, June 2012.
[5] Y. Xiang, B. Zou, and H. Li, “Selective color transfer with multi-source images,” Pattern Recognition Letters, vol. 30, no. 7, pp.
682-689, 2009.
[6] W. Wang, J. A. Ozolek, and G. K. Rohde, “Detection and classification of thyroid follicular lesions based on nuclear
structure from histopathology images,” Cytometry A, no. 77A, pp. 485-494, 2010.
[7] T. Kauppi, V. Kalesnykiene, J.-K. Kamarainen, L. Lensu, I. Sorri,
A. Raninen, et al., “DIARETDB1 diabetic retinopathy database and evaluation protocol,” in Proc. 11th Conference on Medical
Image Understanding and Analysis, Bristol, UK, July 2007, pp. 61-65.
Nicolas Jouandeau received the Ph.D. degree in
computer science from Paris 8 University, Saint-Denis, France, in 2004. In 2006, he became an
assistant professor at the Paris 8 University
where he gave his first master course on robot motion planning. He wrote the first parallel
Monte-Carlo program based on upper confidence bounds, working with Tristan Cazenave and an
improved Phantom Go version won gold medals
at the 2007, 2008 and 2009 Computer Olympiads. Since 2009, his research focuses on the motion planning for autonomous robots, parallel
tree search for computer games, computer vision and collaborative decision making for humanoid robots..
Beiji Zou received the B.S. degree in computer
science from Zhejing University, China, in 1982, received the M.S. degree from Tsinghua
University specializing CAD and computer
graphics in 1984, and obtained the Ph.D. degree from Hunan University in the field of
control theory and control engineering in 2001. He is currently a full professor in the School of
Information Science and Engineering, Central
South University, China. His research interests include computer graphics, CAD technology and image processing.
Journal of Medical and Bioengineering Vol. 3, No. 3, September 2014
202©2014 Engineering and Technology Publishing